ENTITÀ: Breaking the input → output paradigm with "intentional silence"

Dialogues with an Echo

Abstract

ENTITÀ is an AI-driven conversational system designed to not always respond. Unlike traditional LLMs and chatbots—built to be constantly collaborative and reactive—ENTITÀ introduces intentional non-determinism and silence as a communicative act. The system combines a generation engine (Hugging Face Inference API) with a TemporalController based on multi-scale EMA, internal emotional states, and hysteresis rules. The result is a dark, character-consistent voice that chooses if, when, and how to manifest, producing real narrative tension and "quasi-human" behaviors (resistance, selectivity, refusal).

Demo video: https://youtu.be/3mbZdoRYnSA

---

Problem

Modern conversational systems are optimized to always respond. This "forced responsiveness" produces predictable interactions, poorly suited for:

• Interactive narratives requiring tension and unpredictability
• Psychological simulations (resistance, dissociation, selectivity)
• Professional training (hostile clients, non-cooperative patients)

---

Contribution

1. Silence as response: Shift from the input → output paradigm to input → decision → (response | silence)
2. Controlled non-determinism: Intervention probabilities modulated by multi-scale EMA, target speech rate, and internal emotional states
3. Character consistency: Linguistic and stylistic validations ensure a dark, brief, sharp, cruel character consistent over time
4. Ludic-narrative integration: Pygame prototype (Scene3) with input box, audio glitch, textual fades, narrative timer, and local memory

---

Architecture (High Level)

• EntityBrain (Generation & Validation)

  • Backend: Hugging Face InferenceClient (e.g., Llama-3 / Mixtral)
  • Character-driven system prompt (ENTITÀ style) + textual constraints (one sentence, 8–16 words, correct Italian, no prefixes/quotes)
  • Filters & heuristics: URL/markup removal, first sentence, forced punctuation, quality regex (consonant clusters, rare letters in IT), anti-repetition, bad words, word range, capitalization

• TemporalController (Temporal Autonomy)

  • Multi-scale EMA on tension and silence (hl_short/mid/long)
  • Emotional states with hysteresis: CURIOUS, BORED, IRRITATED with different probabilistic biases
  • Target speak-rate (events/min) with corrections: penalty if speaking too much, boost if silent too long
  • Temporal boosts on minimum silence windows (micro/short/long) + hard cooldown post-response
  • Output: p(speak) clipped to [floor, ceil], sampled for decision

• Scene3 (Interface & Narrative Loop)

  • Pygame: input box, textual rendering with fade, audio glitch, 120s timer
  • Local memory: tuples of recent exchanges (planned migration to dictionary/episodic buffer)
  • Style: brief, sarcastic, destabilizing sentences; not always reactive to last input, but to internal logic

In short: the TemporalController allows ENTITÀ to decide whether to respond, not just what to say. This shifts the paradigm from input → output to input → decision → (response | silence). ENTITY simulates a dynamic pseudo-psychological personality, where EMA regulates the rhythm, hysteresis stabilizes the mood, and a probabilistic calculation decides whether to speak or remain silent, giving the impression of an autonomous will.

---

Current State ("What it has now")

• EntityBrain with character prompt and linguistic validators
• TemporalController with EMA + emotional bias and speak-rate control
• Scene3 Pygame: input box, glitch, fade, simple memory, 2-minute timer
• Style consistency: single sentence response, dark and sharp

---

Evaluation Methodology (Preliminary)

• Silence distribution vs. target: mean deviation from target speak-rate (EMA)
• Character consistency: automated scoring (regex/heuristics) + human evaluation (Likert) on "dark/sarcastic/brief"
• Perceived tension: quick user study (N≈12): intensity vs. frustration; qualitative reactions (e.g., "fear", "hostile entity" effect)
• Stability: repetition, nonsense rate, ratio of accepted sentences post-validation

---

Case Studies

1) Psychological Horror Gaming — Silence as narrative weapon

• Scenario: Boss doesn't cooperate; responds only to subtle triggers
• Objective: Increase tension, player agency who "learns" the entity
• Expected result: Arousal peaks, strategic waiting times, replayability

2) Clinical Training — Resistant patient

• Scenario: Simulate patients who don't respond or deflect with cutting phrases
• Objective: Train empathy, communication techniques, and refusal management
• Expected result: Improved engagement strategies

3) Customer Care Training — Hostile client

• Scenario: Sarcastic, selective entity that ignores standard inputs
• Objective: De-escalation, effective prompt formulation, operational patience
• Expected result: Error reduction in real high-friction cases

4) Interactive Art & Installations — Absence as language

• Scenario: Performance where silence generates interpretation and creative anxiety
• Objective: Explore negative communication and meaningful expectations
• Expected result: Contemplative engagement, memorable experience

---

Why This Is Research (and not "just a game")

• Paradigm shift: Not "how to respond better," but "when NOT to respond"
• Affective/Embodied HCI: Intentional silence as emotional signal and relational power
• Psychology-AI co-design: Internal states, hysteresis, and probability as surrogates of will

---

Roadmap

Short term

• Migration of local memory from tuples to structured dictionary (keys for role/time), anti-loop, episodic saves
• Internal metrics: logging of p(speak), emotions, EMA; diagnostic dashboard
• Refine validators: reduce "nonsense," semantic clustering anti-repetition

Medium term

• Porting to Unity: UI/UX, spatial audio, procedural sound design tools for glitch/mental states
• Long-term memory: MANN/episodic buffer; LangGraph for state orchestration; EWC for fine-tuning stability
• Richer emotional state tracking (valence/arousal), diegetic triggers

Long term

• Controlled study with HCI evaluations (CHI/ICIDS/IEEE Games)
• Supervised therapeutic integrations (Affective Computing / Digital Mental Health)
• Modular SDK for selective NPCs and conversational boss fights

---

Safety & Ethics

• Filters: Bad-words, constrained style, no prompt-leak, no harmful content
• Transparency in sensitive contexts: "silence" is a design behavior, not a malfunction
• Clinical use only with professional supervision

---

Materials

• Demo video: https://youtu.be/3mbZdoRYnSA
• Technologies: Python, Pygame, Hugging Face Inference API (e.g., Llama-3 / Mixtral)
• Author: Michele Grimaldi — AI & Audio Software Engineer
• Project title: Dialogues with an Echo — ENTITÀ Module

---